2025 Will Be Make or Break for the Gen AI Bubble
Thus far, the amount of compute has been the most important variable in driving Gen AI model evolution. Therefore, unless one of the Gen AI companies innovates a new method by which to train models, Gen AI model evolution will be purely a scale game. Yet, will it matter? Anthropic threw 4x the compute at its best-in-class Claude 3.5 model, yet it is only marginally better than Claude 3.0.
1st Place: From where I stand if Gen AI is purely a scale game, it gives the advantage to Google (GOOGL) and its Gemini family of models. Google has the balance sheet and AI has been Google’s DNA since inception back in 1998.
2nd Place: I would place Meta (META) a close second with its Llama family of open source models. Personally, I would love for Llama to become the frontier LLM that other LLMs and applications standardize on. Meta has the balance sheet, and I believe that Meta, given its Facebook and Instagram heritage, has AI experience that Microsoft (MSFT) for example, lacks.
In addition, Zuckerberg has the best attitude of the frontier LLM company leaders. Zuckerberg imagines Gen AI as a force multiplier, whereas if you listen to Sam Altman and Dario Amodei (OpenAI and Anthropic founders respectively), there is a God complex that shines through. Founder CEOs that suffer from a God complex tend to over-promise and under-deliver. Larry Ellison (ORCL) is one example. Elon Musk (TSLA) is another.
3rd place: Third place is a tie and goes to OpenAI and Anthropic. While OpenAI kicked off the Gen AI buzz, its GPT4o mini model was cool, but not a game changer. I do not have high hopes for OpenAI’s forthcoming Strawberry model nor do I for its Orion model (2026?) Why?
Because OpenAI lost a bunch of co-founders in recent weeks. Why would they have left if what is coming will be a game-changer?
In addition, OpenAI’s in-process capital round would be for a much higher valuation than the rumored $100 billion-plus figure if the forthcoming models were going to light the world on fire given that OpenAI valued itself at $86 billion back in February.
Further, Anthropic threw 4x the compute at its Claude 3.5 frontier model (perhaps the best-in-class of the frontier models), yet Claude 3.5 is only marginally better than Claude 3.0. Therefore, how will OpenAI’s Strawberry be a game changer? We may have reached the point of diminishing returns as it relates to how Gen AI models scale.
Last, TSMC’s co-founder C.C. Wei believes that Altman’s ambitions are too aggressive.
As for capital requirements, $1 billion is the cost to train the current crop of frontier models, $10 billion for the next gen (2026?) and $100 billion for the following gen (2027?). Who is going to be willing to write a $100 billion check to train an LLM if the two previous generations are underwhelming?